PrePass: Load balancing with data plane resource constraints using commodity SDN switches

Abstract

Software Defined Networks (SDN) enables the network control logic to be designed and operated on a global network view by decoupling the control plane from the data plane. Due to versatility and universality of network applications, network layer load balancing is crucial to ensure operational efficiency in a network with a variety of workloads. However, limited resources (e.g., TCAM and computing capacity) on SDN switches bring critical challenges for load balancing. On one hand, though some solutions satisfy resource constraints, these methods do not work well especially for network asymmetry and traffic dynamics. On the other hand, most previous works achieve load balancing with additional hardware or software resources, which increase the system cost and limit the applicability. Thus, this paper tries to deal with the following challenge: how to achieve load balancing without additional device and software on commodity switches while dealing with network/traffic uncertainties? To this end, we design and implement PrePass, which uses wildcard entries for some aggregate flows to satisfy the flow table size constraint, and performs reactive routing for newly arrived flows to achieve load balancing even with network/traffic uncertainties. We define the problem of load balancing with flow table size constraint, and prove its NP-hardness. We then present an efficient algorithm based on randomized rounding, and analyze that our algorithm can achieve constant bi-criteria approximation under most practical situations. To make our problem more robust, an extended version without traffic size knowledge is also studied. We implement PrePass on a real SDN testbed. The experimental results and extensive simulation results show that our proposed method can satisfy different resource constraints on switches, and only increase the link load ratio by about 5%-10% compared with per-flow routing scheme under various traffic scenarios.